Model

Architecture: ACT with ResNet18 backbone, VAE encoder
Dataset: lerobot/aloha_sim_insertion_human_image (50 episodes, 25k frames)
Training: 100k steps, batch size 8, chunk size 100, lr 1e-5
Final loss: 0.006
Platform: Kaggle 2x Tesla T4

Repository Structure

checkpoints/
└── last/
    ├── pretrained_model/
    │   ├── config.json             # Model architecture config
    │   ├── model.safetensors       # Trained weights
    │   └── train_config.json       # Training hyperparameters
    └── training_state.pt           # Optimizer state (for resume)
aic_policy_wrapper.py               # AIC competition integration wrapper
training_analysis.png               # Visualization of predictions vs ground truth

Usage

from lerobot.policies.act.modeling_act import ACTPolicy
model = ACTPolicy.from_pretrained("farenh/act-aic-cable-insertion", subfolder="checkpoints/last/pretrained_model")
model.eval()

AIC Integration

aic_policy_wrapper.py wraps the trained model for the AIC toolkit Docker container. It converts ROS 2 sensor data (camera images + joint states) into ACT input format and outputs robot motion commands.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support